Re: poor VACUUM performance on large tables

Jan Peterson <jan.l.peterson@gmail.com> writes:
> Based on this, it looks like we could stand to bump up our FSM another
> couple hundred thousand.  Does it buy us anything to reduce the number
> of FSM relations from the default of 1000?

Not a lot; as the comment says, those slots are only about 50 bytes
each.  (I think the true figure is closer to 70, according to some
measurements I did recently on CVS tip, but in any case it's less than
100 bytes apiece.)  Still, a byte saved is a byte earned ...

> It looks like we are using the defaults for work_mem (1024) and
> maintenance_work_mem (16384).  We could certainly bump these up.  Is
> there a good way to determine what settings would be reasonable?

I'd bump up maintenance_work_mem by a factor of 10 and see if it makes a
difference.  It should reduce the number of passes over the indexes when
vacuuming up lots of deleted rows.  If you have lots of RAM you might be
able to increase it more, but try that for starters.

            regards, tom lane